APL activation experiment — prototype

An isolated A/B test: does giving every feedforward neuron its own learnable activation function (Adaptive Piecewise-Linear units — each neuron's own trainable a_s/b_s hinges, not one fixed ReLU shared by the whole network) actually lower held-out loss on this app's own real task suite? Trains two SEPARATE, disposable models ("baseline" = today's architecture, "apl" = the same architecture with learnable per-neuron activations) on identical real data — never touches Language Playground, Chat Model, or Chat Model Lab.